Professional Documents
Culture Documents
Synopsis
Synopsis
ON
Predicting Customer Churn
SUBMITTED IN PARTIAL FULFILLMENT OF THE
REQUIREMENTS
FOR THE AWARD OF THE DEGRE
BACHELOR OF TECHNOLOGY
INFORMATION TECHNOLOGY
Submitted by:
1. Abhay Singh (2005080130001)
2. Ishan Srivastava (2005080130026)
3.Nripendra (2005080130039)
4.Mohammad Talha (2005080130033)
1 Certificate
2 Acknowledgement
3 Introduction
4 Abstract
5 Problem Statement
6 Future Work
7 Limitation
8 Algorithm
Complexity of the
Proposed Model
9 Concluson
Certificate
This is to certify that the report entitled ………… being submitted by Mr/Ms…….. in
partial fulfillment for the award of the Degree of bachelor of Technology in IT to the
Dr. Abdul Kalam Technical University is a record of bonafied work Carried out by
him/her under my guidance and supervision.
The results embodied in this seminar report have not been submitted
to any other University or Institute for the award of any Degree or Diploma.
(Head of the Department) (Project Coordinator)
Acknowledgement
Deep learning (DL) is a new discipline of computer science that extracts patterns from past data and
makes accurate predictions using feature embedding methods. DL has been used successfully in a
variety of felds, including stock price forecasting, personality recognition, disease prediction, text
categorization, and others. Tere is considerable interest in using DL to assist companies in
accurately forecasting customer churn from historical data. As a result, it is critical to conduct
research and use advanced DL approaches to customer data in order to accurately assess customer
churn. As a result, in our research, we present a deep learning technique called BiLSTM-CNN.
BiLSTM retains useful insights from the given Telco dataset context information in both the
forward and backward directions5 . Our major goal is to efectively categorize the data as churn or
non-churn using a BiLSTM-CNN model. Deep learning is used in this model to give more accurate
and efcient churn prediction.
Introduction
Customers are critical to any company's success, thus every effort is made to assure their
satisfaction. Maintaining the happiness of existing consumers is critical in subscription-based
product expansion. The telecommunications sector is very competitive, with multiple suppliers
providing comparable services. A single bad encounter can result in the irreversible loss of a
consumer. Furthermore, a large migration of unsatisfied consumers could have serious financial and
reputational consequences for the organization
Customer churn, a phenomenon that causes large financial losses when customers leave a
business, makes it difficult for modern organizations to retain customers. When
dissatisfied customers find their present company's services inadequate, they frequently
migrate to another service provider. Machine learning and deep learning (ML/DL)
approaches have already been used to successfully identify customer churn. In some
circumstances, however, ML/DL-based algorithms lacks in delivering promising results
for detecting client churn. Previous research on estimating customer churn revealed
unexpected forecasts when utilizing machine learning classifiers and traditional feature
encoding methodologies. Deep neural networks were also used in these efforts to extract
features without taking into account the sequence information. In view of these issues,
the current study provides an effective method for predicting customer churn based on a
hybrid deep learning model termed BiLSTM-CNN. The goal is to effectively estimate
customer churn using benchmark data and increase the churn prediction process's
accuracy. The experimental results show that when trained, tested, and validated on the
benchmark dataset, the proposed BiLSTM-CNN model attained a remarkable accuracy
of 81%.
Problem Statement
The churn identification task is a binary classification that seeks to separate churners from non-
churners. The training data is denoted by P = {p1, p2, p3,. pm}, and the associated class label is
denoted by y, where y∈{0,1}. If y equals one, it denotes churn; otherwise, it represents non-churn.
Our goal is to create a model that can accurately distinguish churn and non-churn based on the
training data and class labels we provided.
Future Work
The following study directions are suggested for the future:
(ii) Future research should look into multidimensional CNN approaches and other
deep neural network models for predicting customer churn
(iii) In future research, the suggested BiLSTM-CNN model should be used to tackle
multiclass issues, multi-label classification, and imbalanced classification.
Limitation
This study has certain limitations, which are listed below:
(ii) The research will only use the BiLSTM-CNN model, without experimenting with
other composite Deep Learning models;
(iv) The study will only use random feature embedding, without considering alternative
representation models such as "pre-trained
(v) Given the dataset contains 20 features, more than twenty features should be used
Time Complexity:
The number of operations executed during the training and prediction phases determines the
time complexity of a BiLSTM-CNN model for churn prediction. The LSTM and CNN layers are
the most significant contributors to time complexity. LSTM Time Complexity: A single LSTM cell
has a time complexity of O(n), where n is the number of input features to the LSTM cell. This
complexity is twice for a BiLSTM layer since it processes input in both forward and backward
directions. The LSTM layer's overall time complexity is O(T * n), where T is the number of time
steps (sequence length) in the input data. CNN Time Complexity: The time complexity of a 1D
Convolutional Neural Network (CNN) layer is determined by the kernel size, the number of filters
(channels), and the length of the input sequence. Convolution takes O(k * n) operations for each
filter, where k is the kernel size. The overall time complexity of the CNN layer is O(F * k * n) if
there are F filters. Overall Time Complexity: The total temporal complexity of the BiLSTM-CNN
model is the sum of the complexities of the BiLSTM and CNN layers, which can be written as: O(T
* n) + O(F * k * n).
Space Complexity:
The number of parameters in the BiLSTM-CNN model, which includes weights, biases, and other
trainable parameters, determines the space complexity of the model. Space Complexity of the
BiLSTM layer is O(L * (4 * n * n + 4 * n)), where L is the number of LSTM cells (units) in the
layer. CNN layer has a space complexity of O(F * (k * n + 1)), where F is the number of filters, k is
the size of the kernel, and n is the number of input features. The BiLSTM-CNN model's overall
space complexity is the sum of the space complexities of the BiLSTM layer and the CNN layer,
which can be expressed as: O(L * (4 * n * n + 4 * n)) + O(F * (k * n + 1)).
Literature review
A literature review on Customer Churn Prediction is included in this section. A review of selected
studies Te study conducted by6 investigates staf attrition through the use of several machine
learning models in order to improve customer satisfaction and retention. Using the IBM dataset, fve
basic models and three ensembles were created and analyzed. Te linear model outperformed the
others in terms of accuracy, recall, and AUC. Te research conducted by7 employs big data analysis
to develop an estimating model for customer attrition in communication frms. For modelling,
segmentation and regression approaches are used with good results. However, more system
enhancements are required employs big data analysis to develop an estimating model for customer
attrition in communication frms. For modelling, segmentation and regression approaches are used
with good results. However, more system enhancements are required. Ranjan and Sood8
investigated the application of Twitter sentiment analysis to forecast customer attrition in Indian
telecommunications. For prediction, they used the Nave Bayes classifer and TextBlob, evaluated
the models with IBM SPSS, and discovered positive results for increasing customer experience and
retention. However, the restricted dataset need further.
Expansion for more robust results. Jeyakarthic et al.9 developed an ML-based customer churn
prediction model in a cloud computing setting. With 95.50 precision, 70.49 recall, 91.71 accuracy,
95.13 F-score, and 67.20 kappa value, the model performed well. Te study advises that feature
selection and clustering approaches be used to improve the model further. Ahmad et al.10 used
machine learning techniques on large amounts of data to create a client attrition prediction model
for the telecom industry. Te decision tree, random forest, gradientboosted machine tree, and
extreme gradient-boosted machine tree techniques were all used in the model. Te XGBOOST
algorithm performed the best among them. Panjasuchat et al.11 used supervised learning datasets to
implement reinforcement learning for customer churn prediction. When the data amount was
increased, DQN beat XGBoost, Random Forest, and KNN. However, when the dataset pattern
changed, the performance of all methods declined. Nguyen et al.12 investigated customer attrition
in service industries and dealt with data imbalance issues. Tey contrasted SMOTE and Deep Belief
Network with cost-sensitive data resampling approaches, weighted loss, and focal loss. In low
turnover rate conditions, focal loss and weighted loss surpassed SMOTE and DBN in prediction
performance. Wahul et al.13 used SGD, RF, GB, AdaBoost, and Stacking classifers to create an
ensemble learning architecture for churn prediction. Te stacked model outperformed individual
classifers in identifying churn consumers due to better accuracy, recall, and AUC. Te researchers
recommend experimenting with advanced ensemble approaches and diverse data sources. Prabadevi
et al.14 used nine months of customer data to examine machine-learning algorithms for early
customer attrition prediction. In terms of accuracy, the Stochastic Gradient Booster surpassed other
methods. For hyperparameter tweaking, the study recommends employing more complex
optimization approaches. Torat et al.15 investigated the efectiveness of deep learning in forecasting
customer attrition in the telecom business. Algorithms such as Random Forest and XGBoost were
used in the study. Te deep learning model deployed achieved 88% accuracy, although more data
and hyperparameter optimization could improve outcomes. Saha et al.16 evaluated multiple
learning approaches, including CNN and ANN, using two public datasets to construct a churn
prediction model. On the frst dataset, CNN obtained 99% accuracy and 98% on the second. For
better prediction, the study proposed utilizing structured, unstructured, and behavioral data. Seymen
et al.17 developed ANN and CNN models for predicting retail customer attrition and compared
them to various machine learning algorithms. Te deep learning-based CNN model beat the others,
reaching 97.62% classifcation accuracy. Te study advises employing AI technologies to investigate
missing client behavior patterns. Research gap and justifcation for using BiLSTM-CNN model for
churn prediction While basic machine and deep learning techniques have shown efcacy in customer
churn prediction, earlier research have struggled to achieve greater classifcation accuracy levels.
Incorrect parameter and layer selection can have a major impact on neural network model
performance. Te suggested BiLSTM+CNN model will investigate a variety of layers and parameter
values to solve customer attrition in the telecoms industry. In addition, we will run further deep
learning model iterations and compare their outcomes to earlier research. Te proposed approach,
which combines bidirectional long-term short-term memory (BiLSTM) with multiplelayer
convolutional neural networks (CNN), tries to efectively identify customer turnover using
accessible data. Here are some of the reasons why the suggested BiLSTM-CNN architecture is
appropriate for churn prediction: • Bidirectional LSTM: A bidirectional LSTM has two LSTM
layers: one that processes the input sequence forward and one that processes the input sequence
backward. Tis can help the model perform better on tasks where the order of the input sequence is
essential. Te order of the input sequence is signifcant in churn prediction because it can reveal
patterns that suggest whether a client is likely to churn. For example, if a customer has lately made
a big number of transactions, it could signal that they are happy with the service and are less likely
to churn. However, if a customer has recently cancelled their service, it may suggest that they are
dissatisfed with the service and are and are more likely to churn12. • Convolutional Neural
Network: A convolutional neural network (CNN) is a sort of neural network that works well with
sequential data. CNNs can learn to extract features from input sequences and utilize them to
produce predictions. CNNs can be used in churn prediction to extract information from a
customer’s past data, such as their purchase history, service usage, and interactions with customer
care. Tese characteristics can then be used to forecast if a customer is likely to churn18. In addition
to the benefts listed above, the following are some additional advantages of employing BiLSTM-
CNN for churn prediction: • It can detect long-term dependencies in the input sequence: BiLSTM-
CNN can detect long-term dependencies in the input sequence. Tis is signifcant for predicting
customer turnover since customer churn is frequently driven by a set of events that occur over time.
For example, a client may be unsatisfed with the service for some time before deciding to churn.
Tese long-term dependencies can be captured by BiLSTM-CNN and used to produce more accurate
predictions (20). It is relatively easy to train: BiLSTM-CNN is relatively easy to train compared to
other deep learning models. Tis is because BiLSTM-CNN has a relatively simple architecture19.
Methodology
We suggested technique is divided into following modules (Fig. 1): (i) Churn dataset acquisition, (ii) preprocessing, (iii)
Applying deep learning model, and (iv) Applied example for Customer churn prediction. . Each module’s specifcs are
provided below.
In conclusion, building an effective neural network model in R using Keras for predicting
customer churn is a valuable asset for businesses. By leveraging advanced analytics, companies
can proactively address customer churn and enhance customer retention strategies